
The Complexity of Gradient Descent: CLS = PPAD \ PLS John Fearnley Paul W. Goldberg University of Liverpool, United Kingdom University of Oxford, United Kingdom [email protected] [email protected] Alexandros Hollender Rahul Savani University of Oxford, United Kingdom University of Liverpool, United Kingdom [email protected] [email protected] Abstract We study search problems that can be solved by performing Gradient Descent on a bounded convex polytopal domain and show that this class is equal to the intersection of two well-known classes: PPAD and PLS. As our main underlying technical contribution, we show that computing a Karush-Kuhn-Tucker (KKT) point of a continuously differentiable function over the domain [0; 1]2 is PPAD \ PLS-complete. This is the first natural problem to be shown complete for this class. Our results also imply that the class CLS (Continuous Local Search) { which was defined by Daskalakis and Papadimitriou as a more \natural" counterpart to PPAD \ PLS and contains many interesting problems { is itself equal to PPAD \ PLS. arXiv:2011.01929v3 [cs.CC] 13 Apr 2021 1 1 Introduction It is hard to overstate the importance of Gradient Descent. As noted by Jin et al.[2021], \Machine learning algorithms generally arise via formulations as optimization problems, and, despite a massive classical toolbox of sophisticated optimization algorithms and a major modern effort to further develop that toolbox, the simplest algorithms|gradient descent, which dates to the 1840s [Cauchy, 1847] and stochastic gradient descent, which dates to the 1950s [Robbins and Monro, 1951]|reign supreme in machine learning." Jin et al.[2021] continue by highlighting the simplicity of Gradient Descent as a key selling-point, and the importance of theoretical analysis in understanding its efficacy in non-convex optimization. In its simplest form, which we consider in this paper, Gradient Descent attempts to find a minimum of a continuously differentiable function f over some domain D, by starting at some point x0 and iterating according to the update rule xk+1 xk − ηrf(xk) where η is some fixed step size. The algorithm is based on the fundamental fact that for any point x the term −∇f(x) points in the direction of steepest descent in some sufficiently small neighbourhood of x. However, in the unconstrained setting|where the domain is the whole space|it is easy to see that Gradient Descent can at best find a stationary point. Indeed, if the gradient is zero at some point, then there is no escape. Note that a stationary point might be a local minimum, but it could also be a saddle point or even a local maximum. Similarly, in the constrained setting|where the domain D is no longer the whole space|Gradient Descent can at best find a point x that satisfies the Karush-Kuhn-Tucker (KKT) optimality conditions. Roughly, the KKT conditions say that the gradient of f is zero at x, or if not, x is on the boundary of D and any further local improvement would take us outside D. In this paper we investigate the complexity of finding a point where Gradient Descent terminates|or equivalently, as we will see, a KKT point|when the domain is bounded. It is known that a global or even a local minimum cannot be found in polynomial time unless P = NP [Murty and Kabadi, 1987; Ahmadi and Zhang, 2020a]. Indeed, even deciding whether a point is a local minimum is already co-NP-hard [Murty and Kabadi, 1987]. In contrast, it is easy to check whether a point satisfies the KKT conditions. In general, finding a KKT point is hard, since even deciding whether a KKT point exists is NP-hard in the unconstrained setting [Ahmadi and Zhang, 2020b]. However, when the domain is bounded, a KKT point is guaranteed to exist! This means that in our case, we are looking for something that can be verified efficiently and that necessarily exists. Intuitively, it seems that this problem should be more tractable. This intuition can be made formal by noting that these two properties place the problem in the complexity class TFNP of total search problems in NP: any instance has at least one solution, and a solution can be checked in polynomial time. A key feature of such problems is that they cannot be NP-hard unless NP = co-NP [Megiddo and Papadimitriou, 1991]. TFNP problems have been classified via certain \syntactic subclasses" of TFNP, of which PPAD and PLS are two of the most important ones. 1.1 NP total search classes: PPAD, PLS, and CLS As discussed by Papadimitriou[1994], TFNP is unlikely to have complete problems, and various syntactic subclasses have been used to classify the many diverse problems that belong to it. Among them, the classes PPAD and PLS (introduced by Papadimitriou[1994] and Johnson et al.[1988] respectively) have been hugely successful in this regard. Each of 2 these classes has a corresponding computationally inefficient existence proof principle, one that when applied in a general context, does not yield a polynomial-time algorithm.1 In the case of PPAD this is the parity argument on a directed graph, equivalent to the existence guarantee of Brouwer fixpoints: a Brouwer function is a continuous function f : D ! D where D is a convex compact domain, and Brouwer's fixed point theorem guarantees a point x for which f(x) = x. PPAD has been widely used to classify problems of computing game-theoretic equilibria (a long line of work on Nash equilibrium computation beginning with Daskalakis et al.[2009]; Chen et al.[2009b], and market equilibria, e.g., Chen et al. [2009a]). PPAD also captures diverse problems in combinatorics and cooperative game theory [Kintali et al., 2013]. PLS, for \Polynomial Local Search", captures problems of finding a local minimum of an objective function f, in contexts where any candidate solution x has a local neighbourhood within which we can readily check for the existence of some other point having a lower value of f. Many diverse local optimization problems have been shown complete for PLS, attesting to its importance. Examples include searching for a local optimum of the TSP according to the Lin-Kernighan heuristic [Papadimitriou, 1992], and finding pure Nash equilibria in many-player congestion games [Fabrikant et al., 2004]. The complexity class CLS (\Continuous Local Search") was introduced by Daskalakis and Papadimitriou[2011] to classify various important problems that lie in both PPAD and PLS. PPAD and PLS are believed to be strictly incomparable|one is not a subset of the other|a belief supported by oracle separations [Beame et al., 1998]. It follows from this that problems belonging to both classes cannot be complete for either one of them. CLS is seen as a strong candidate for capturing the complexity of some of those important problems, but, prior to this work, only two problems related to general versions of Banach's fixed point theorem were known to be CLS-complete [Daskalakis et al., 2018; Fearnley et al., 2017]. An important result|supporting the claim that CLS-complete problems are hard to solve|is that the hardness of CLS can be based on the cryptographic assumption of indistinguishability obfuscation [Hub´aˇcekand Yogev, 2017]. Prior to the present paper, it was generally believed that CLS is a proper subset of PPAD \ PLS, as conjectured by Daskalakis and Papadimitriou[2011]. 1.2 Our contribution and its significance Our main result is to show that finding a point where Gradient Descent on a continuously differentiable function terminates|or equivalently a KKT point|is PPAD \ PLS-complete, when the domain is a bounded convex polytope. This continues to hold even when the domain is as simple as the unit square [0; 1]2. The PPAD \ PLS-completeness result applies to the \white box" model, where functions are represented as arithmetic circuits. Computational Hardness. As an immediate consequence, our result provides convincing evidence that the problem is computationally hard. First of all, there are reasons to believe that PPAD \ PLS is hard simply because PPAD and PLS are believed to be hard. Indeed, if PPAD \ PLS could be solved in polynomial time, then, given an instance of a PPAD- complete problem and an instance of a PLS-complete problem, we would be able to solve at least one of the two instances in polynomial time. Furthermore, since CLS ⊆ PPAD \ PLS, the above-mentioned cryptographic hardness of CLS applies automatically to PPAD \ PLS, and thus to our problem of interest. 1The other well-known such classes, less relevant to the present paper, are PPA and PPP; it is known that PPAD is a subset of PPA and also of PPP. These set-theoretic containments correspond directly to the strength, or generality, of the corresponding proof principles. 3 Continuous Local Search. Since Gradient Descent is just a special case of continuous local search, our hardness result implies that CLS = PPAD \ PLS which disproves the widely believed conjecture by Daskalakis and Papadimitriou[2011] that the containment is strict. Our result also allows us to resolve an ambiguity in the original definition of CLS by showing that the high-dimensional version of the class reduces to the 2-dimensional version of the class (the 1-dimensional version is computationally tractable, so no further progress is to be made). Equality to PPAD \ PLS also applies to a linear version of CLS analogous to the class Linear-FIXP of Etessami et al.[2020]. PPAD \ PLS. Perhaps more importantly, our result establishes PPAD \ PLS as an important complexity class that captures the complexity of interesting problems. It was previously known that one can construct a problem complete for PPAD \ PLS by gluing together two problems, one for each class (see Section 2.2), but the resulting problem is highly artificial.
Details
-
File Typepdf
-
Upload Time-
-
Content LanguagesEnglish
-
Upload UserAnonymous/Not logged-in
-
File Pages80 Page
-
File Size-